Learning-based filtering algorithm to recognize and retrieve features of swine lagoons in agricultural landscapes using aerial ortho-imagery

Authors: Sohoulande Djebou, Dagbegnon; WANG, SHUO - University Of Texas At Arlington; Martin, Jerry; JOBE, ROBERT - University Of Texas At Arlington; Ro, Kyoung; Vanotti, Matias; Ducey, Thomas; Birru, Girma

Submitted to: Smart Agricultural Technology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/20/2025
Publication Date: 8/21/2025
Citation: Sohoulande Djebou, D.C., Wang, S., Martin, J.H., Jobe, R., Ro, K.S., Vanotti, M.B., Ducey, T.F., Birru, G.A. 2025. Learning-based filtering algorithm to recognize and retrieve features of swine lagoons in agricultural landscapes using aerial ortho-imagery. Smart Agricultural Technology. https://doi.org/10.1016/j.atech.2025.101347.
DOI: https://doi.org/10.1016/j.atech.2025.101347

Interpretive Summary: In the intensive swine production regions of the Southeast United States (US), the design of swine farms is very distinctive as it includes concentrated feeding barns and open-air lagoons where swine waste is collected and stored for anaerobic decay of organic matter. These open-air lagoons are sources of gas emissions, and they can impact the environment during flooding events. Because swine farms can vary significantly in footprints, a thorough estimate of their geographical coordinates and sizes could be useful to trace and upscale their potential effects on water, and air quality at county, watershed, and State levels. Therefore, this study developed a filtering algorithm for detecting and retrieving the physical features of swine waste treatment lagoons from high-resolution aerial images. Indeed, the lagoon-based swine farms are recognizable from the visible aerial images, because of the rectangular and metallic structure of the concentrated feeding barns, and the regular shapes of the open-air lagoons which are situated in the vicinity of the barns. The developed algorithm uses feature extraction techniques to retrieve barns and waterbodies from Red-Green-Blue (RGB) and near-infrared (NIR) images. Thresholds of normalized differential water index (NDWI) were used to optimize the feature extraction. The filtering algorithm was separately calibrated and validated using the US National Agriculture Imagery Program (NAIP) ortho-imagery over two agricultural watersheds in North Carolina. The algorithm’s accuracy at detecting lagoon-based swine farms was above 98%. In addition, the algorithm estimates of lagoon areas were acceptable. The algorithm’s performance sustained its potential use for surveying swine waste treatment lagoons and upscaling their potential effects on water and air quality.

Technical Abstract: In the intensive swine production regions of the Southeast United States (US), the design of swine farms is very distinctive as it includes concentrated feeding barns and open-air lagoons where swine waste is collected and stored for anaerobic decay of organic matter. These open-air lagoons are sources of harmful gas emissions, and they can cause environmental pollution during flooding events. To compound these issues, swine farms can vary significantly in footprint, can be found in high densities in constrained geographical areas, and their distribution in the agricultural landscapes is often erratic. Given these conditions, a thorough estimate of anaerobic swine lagoon geolocations, areas, and shapes could be useful to trace and upscale the potential effects of the swine waste management systems on water, and air quality at county, watershed, and State levels. Therefore, this study developed a filtering algorithm for detecting and retrieving the physical features of swine waste treatment lagoons from high-resolution aerial images. Indeed, the lagoon-based swine farms are recognizable from the visible spectrum of aerial images, because of the rectangular and metallic structure of the concentrated feeding barns, and the regular shapes of the open-air lagoons which are situated in the vicinity of the barns. The developed algorithm uses feature extraction techniques to retrieve signals of barns and waterbodies from Red-Green-Blue (RGB) and near-infrared (NIR) bands of the electromagnetic spectrum. Thresholds of normalized differential water index (NDWI) were used to optimize the feature extraction. The filtering algorithm was separately calibrated and validated using the US National Agriculture Imagery Program (NAIP) ortho-imagery over two agricultural watersheds in North Carolina. The algorithm’s accuracy at detecting lagoon-based swine farms was above 98% with Fscore values above 82%. In addition, the algorithm estimates of lagoon areas were acceptable with R2 above 0.88 and Nash Sutcliffe efficiency (NSE) above 0.81. Hence, the algorithm’s performance sustained its potential use for surveying swine waste treatment lagoons and upscale their potential effects on environmental quality.